Fortinet adds AI prompt inspection to existing FortiGate firewalls through a cloud service

Fortinet is promising AI security without another system for customers to deploy. On October 26, 2026, it announced FortiGuard Inline AI Security Service, bringing cloud-delivered inspection to existing FortiGate firewalls. The change goes beyond spotting AI use: Fortinet says the service can inspect interactions and mask sensitive data as it passes through.

From spotting AI use to reading it

FortiGate firewalls could already identify shadow AI—AI use outside an organization’s approved systems—and apply rules to prevent data loss, SiliconANGLE reported. The new service adds inspection of the content inside those exchanges, rather than stopping at identifying the service or tracking sensitive information moving across the network.

The inspection technology comes from FortiAIGate, which Fortinet also offers as a gateway between AI applications and models. The cloud service uses its guardrails to check prompts and data in real time across interactions with language models, agents and Model Context Protocol tools. Those tools connect AI systems with other software and data.

Fortinet’s stated concern is that harmful instructions can hide in prompts, model responses and tool communications, then trigger risky tool calls or agent actions. It says FortiAIGate evaluates interactions using the model, application, user role and business context—not only the network addresses or signatures examined by traditional security tools.

A gateway for budgets, permissions and audit records

Alongside the service launch, Fortinet detailed broader FortiAIGate capabilities. That distinction matters: the announcement describes both a cloud-delivered route for firewall customers and a gateway with its own governance functions. The gateway’s inspection covers text, images, audio, video and code, extending beyond typed exchanges with an assistant.

Fortinet says expanded data-loss prevention detects 35 types of personal and confidential information. Administrators can choose whether to receive an alert, deny the exchange or redact the information before exposure. These are configurable responses, rather than a single blanket rule that blocks every interaction containing sensitive material.

The associated controls also address who can use models, how much they can spend and what administrators can review afterward:

• Spending limits: Administrators can assign model permissions, budgets and rate limits by organization, workspace, project or user. Alerts warn as spending approaches a limit; requests are blocked once spending exceeds it.
• Team boundaries: Dedicated workspaces give business units their own policies, models, logs, roles and permissions while retaining centralized oversight.
• Audit evidence: More than 50 predefined policies cover governance, privacy, regulation and security frameworks. Policies can be customized, and logs record prompts, outputs and tool calls for auditors.

Fortinet also says FortiAIGate blocks malicious prompts designed to inflate computing bills before they consume tokens or GPU resources. Request routing can select a model based on content, policy and cost without redesigning the application. A separate visibility function shows which agents are operating and their connections to applications, tools, data and servers.

Local processing remains a separate choice

Fortinet positions direct FortiAIGate deployment for organizations with more demanding security, privacy or data-sovereignty requirements. That option runs on Kubernetes with Nvidia GPUs, either on-premises or in a customer’s cloud. The company says prompts, responses, data and logs stay within the customer’s environment, giving it control over where processing happens.

Some of the underlying technology comes from Virtue AI, whose acquisition Fortinet announced on August 17. At the time, Fortinet described the business as providing runtime protection, automated AI validation and security for autonomous systems. The October announcement explicitly links technology from that acquisition to FortiAIGate’s continuing development.

The announcement leaves performance harder to assess. Fortinet describes its inspection as low-latency with minimal false positives, but supplies no numerical results for either measure. Its concrete deployment promise is narrower: FortiGate customers can add cloud-delivered inspection without managing separate AI-security infrastructure. That does not make it the same processing arrangement as a self-hosted gateway.